RL evals have a cost problem: expensive tests cap your sample sizes

jessi_cata · x · 2026-11-21

jessicata points out an often-overlooked constraint in RL training: the cost of the tests themselves. Some capabilities are verifiable — e.g. checking whether an AI completed a pre-training run at a given cross-entropy — but such verifications can be extremely expensive, directly limiting feasible RL sample sizes.

Related event: Researcher Warns Evaluation Cost Limits RL Sample Sizes(2 posts)→

Original post →

More from Research

Research channel →